simplefuzzyset 0.0.12

Adding to the data structure

First let’s look at adding a string, ‘michaelich’ to an empty set. We first break apart the string into n-grams (strings of length n). So trigrams of ‘michaelich’ would look like:

'-mi'
'mic'
'ich'
'cha'
'hae'
'ael'
'eli'
'lic'
'ich'
'ch-'

Note that fuzzyset will first normalize the string by removing non word characters except for spaces and commas and force everything to be lowercase.

Next the fuzzyset essentially creates a reverse index on those grams. Maintaining a dictionary that says:

'mic' -> (1, 0)
'ich' -> (2, 0)
...

And there’s a list that looks like:

[(3.31, 'michaelich')]

Note that we maintain this reverse index for all grams from gram_size_lower to gram_size_upper in the constructor. This becomes important in a second.

Retrieving

To search the data structure, we take the n-grams of the query string and perform a reverse index look up. To illustrate, let’s consider looking up 'michael' in our fictitious set containing 'michaelich' where the gram_size_upper and gram_size_lower parameters are default (3 and 2 respectively).

We begin by considering first all trigrams (the value of gram_size_upper). Those grams are:

'-mi'
'mic'
'ich'
'cha'
'el-'

Then we create a list of any element in the set that has at least one occurrence of a trigram listed above. Note that this is just a dictionary lookup 5 times. For each of these matched elements, we compute the cosine similarity between each element and the query string. We then sort to get the most similar matched elements.

If use_levenshtein is false, then we return all top matched elements with the same cosine similarity.

If use_levenshtein is true, then we truncate the possible search space to 50, compute a score based on the levenshtein distance (so that we handle transpositions), and return based on that.

In the event that none of the trigrams matched, we try the whole thing again with bigrams (note though that if there are no matches, the failure to match will be quick). Bigram searching will always be slower because there will be a much larger set to order.